More scalable QTP prototype #13505
More scalable QTP prototype #13505
Conversation
| @ManagedObject("A striped thread pool") | ||
| public class StripedQueuedThreadPool extends ContainerLifeCycle implements ThreadFactory, SizedThreadPool, Dumpable, TryExecutor, VirtualThreads.Configurable | ||
| { | ||
| private static final int STRIPES = 16; |
There was a problem hiding this comment.
Should be configurable, perhaps default to the number of cores.
There was a problem hiding this comment.
Also need to reduce striping for small pools. If max threads is 100, then slicing 16 ways is too much.
| } | ||
|
|
||
| @Override | ||
| public void setMaxThreads(int threads) |
There was a problem hiding this comment.
Add a check that threads should be an exact multiple of STRIPES?
| } | ||
|
|
||
| @Override | ||
| public void setMinThreads(int threads) |
There was a problem hiding this comment.
Add a check that threads should be an exact multiple of STRIPES?
Also, threads should be divided by STRIPES.
| @Override | ||
| public int getMinThreads() | ||
| { | ||
| return queuedThreadPools[0].getMinThreads(); |
| throw new IllegalArgumentException("max threads (" + maxThreads + ") less than min threads (" + minThreads + ")"); | ||
| for (int i = 0; i < queuedThreadPools.length; i++) | ||
| { | ||
| queuedThreadPools[i] = new QueuedThreadPool(); |
| @Override | ||
| protected void doStart() throws Exception | ||
| { | ||
| super.doStart(); | ||
| for (QueuedThreadPool queuedThreadPool : queuedThreadPools) | ||
| { | ||
| queuedThreadPool.start(); | ||
| } | ||
| } | ||
|
|
||
| @Override | ||
| protected void doStop() throws Exception | ||
| { | ||
| super.doStop(); | ||
| for (QueuedThreadPool queuedThreadPool : queuedThreadPools) | ||
| { | ||
| queuedThreadPool.stop(); | ||
| } | ||
| } |
There was a problem hiding this comment.
No need for this if the striped QTP are added as beans, which is also better for dump purposes.
| setIdleTimeout(60000); | ||
| setStopTimeout(5000); | ||
| setReservedThreads(-1); | ||
| // setThreadPoolBudget(new ThreadPoolBudget(this)); |
There was a problem hiding this comment.
The whole ThreadPoolBudget implementation needs to be reworked, we should first confirm we want to go with this solution.
lorban
force-pushed
the
experiment/jetty-12.1.x/AES-perf-investigations
branch
from
8a5ccb6 to
4d517ca
Compare
| @ManagedObject("A striped thread pool") | ||
| public class StripedQueuedThreadPool extends ContainerLifeCycle implements ThreadFactory, SizedThreadPool, Dumpable, TryExecutor, VirtualThreads.Configurable | ||
| { | ||
| private static final int STRIPES = 16; |
There was a problem hiding this comment.
Also need to reduce striping for small pools. If max threads is 100, then slicing 16 ways is too much.
Striping the queue would have my preference, and the fact that we don't need a queue that strictly enforces FIFO (despite the fact that it would break the Maybe we could give the conversant disruptor or the JCTools MPMC |
|
Despite the shortcomings, a striped QTP is still our best option IMHO. |
Signed-off-by: Ludovic Orban <[email protected]>
lorban
force-pushed
the
experiment/jetty-12.1.x/AES-perf-investigations
branch
from
4d517ca to
6800c1a
Compare
…dPool Signed-off-by: Ludovic Orban <[email protected]>
lorban
force-pushed
the
experiment/jetty-12.1.x/AES-perf-investigations
branch
from
6800c1a to
10cfdcd
Compare
|
I tried several variations of the jctools queue in the QTP JMH benchmark and all were slower (30k - 100k ops/s) than BAQ (300k ops/s). |
Signed-off-by: Ludovic Orban <[email protected]>
lorban
force-pushed
the
experiment/jetty-12.1.x/AES-perf-investigations
branch
from
ea58469 to
e3c7b04
Compare
Prototype
QueuedThreadPoolalternative that stripes execution requests.